acceptodds
Withdrawn from ICLR 2027

LingGS: Decoupling Feed-Forward Geometry Prediction from Camera Tracking for High-Fidelity Gaussian SLAM

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Achieving high-quality reconstruction and responsive camera tracking remains challenging for monocular RGB Gaussian splatting SLAM under a limited optimization budget. We present an online system that addresses this challenge by combining learned geometric priors with explicit geometric camera tracking, which can substantially outperform learned camera prediction in terms of tracking accuracy. A frozen streaming 3D foundation model serves as a virtual RGB-D sensor, providing depth estimation from monocular video. An explicit geometric solver then estimates camera motion through feature reprojection, and a streaming Gaussian mapper integrates the resulting observations into a dense scene representation. We further show that geometric tracking improves camera pose accuracy across representative streaming geometry backbones, demonstrating the broader applicability of our approach. Extensive experiments on indoor and outdoor datasets show that our method significantly improves reconstruction quality over existing approaches, achieving 7 FPS on a single RTX 4090 and delivering 2–3× the throughput of the evaluated prior SLAM systems.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Reconstruction, Gaussian Splatting SLAM, Camera tracking, Feedforward

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_NNE5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents LingGS, an online monocular Gaussian SLAM system that combines a frozen LingBot-Map geometry predictor, an ORB-SLAM3-based tracker, and an OTF-style incremental Gaussian mapper. Predicted depth is treated as a virtual RGB-D observation, while camera poses are estimated through feature reprojection and a direct depth residual with constant relative uncertainty. The mapper uses Laplacian-guided Gaussian insertion and limited refinement with RGB/depth supervision and keyframe replay. Experiments on four indoor/outdoor datasets report improved rendering quality, lower rendered-depth errors.

Critical strengths
  1. The method achieves the best reported novel-view synthesis metrics across four datasets and the lowest rendered-depth error across three datasets, demonstrating practical potential for monocular Gaussian SLAM.
  2. The ablations show that explicit geometric tracking improves pose accuracy over native streaming-model camera heads, with preliminary evidence across four backbones.
Critical weaknesses
  1. The method primarily combines ORB-SLAM3, LingBot-Map, and On-the-Fly NVS, with the main modification being a direct-depth residual and a manually specified uncertainty model. The virtual-stereo variant already achieves the same PSNR (26.35 dB), while ATE improves only from 79.5 cm to 73.8 cm. The ablation does not isolate the effects of residual parameterization and uncertainty weighting, leaving the contribution beyond existing components insufficiently established.

  2. The evaluation omits relevant recent methods, particularly ARTDECO, as well as VGGT-SLAM, VGGT-SLAM 2.0, and MASt3R-SLAM on shared tasks. Moreover, held-out-view evaluation and depth-scale alignment protocols lack sufficient detail, and no continuous reconstruction video is provided to assess temporal consistency and online map evolution.

  3. The method relies on ORB feature matching and does not address loop closure or dynamic scenes, as acknowledged by the authors. Its robustness in repetitive, weakly textured, or long-term sequences remains unclear, limiting the strength of its claims for general online reconstruction.

Questions for authors
  1. What is the principal methodological contribution beyond combining existing components? Can the authors isolate the effects of residual parameterization and uncertainty weighting, and justify the fixed 5% relative-depth uncertainty?

  2. Why is ARTDECO excluded despite its close relevance? Can the authors compare against ARTDECO for reconstruction quality and efficiency, and evaluate VGGT-SLAM, VGGT-SLAM 2.0, and MASt3R-SLAM on shared tracking and geometric reconstruction tasks?

  3. How are held-out views selected, and how are their camera poses obtained? Is depth-scale alignment performed per sequence or per view, and is the same protocol applied to all methods?

  4. When local bundle adjustment revises historical poses, how are these corrections propagated to the existing Gaussian map? Does local BA use the proposed direct-depth residual or retain the original virtual-stereo formulation?

  5. Can the authors provide a continuous reconstruction demonstration and a runtime breakdown for geometry prediction, tracking, and mapping to substantiate the claims of high-fidelity online reconstruction?

Reviewer_TX5yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a 3D Gaussian Splatting SLAM framework for monocular RGB video that mitigates camera pose estimation errors in streaming 3D foundation models. The main idea is to decouple feed-forward geometry prediction from camera tracking by using a frozen streaming 3D foundation model as a virtual RGB-D sensor, retaining only its initial intrinsics and predicted depths while discarding its per-frame predicted poses, and estimating poses via an ORB-SLAM3-based feature front-end modified with a direct depth residual. For mapping, 3D Gaussians are incrementally inserted from predicted dense depth maps and refined over keyframes. Quantitative evaluations across four indoor and outdoor datasets show that the proposed method outperforms MonoGS, S3PO-GS, Photo-SLAM, and OTF in rendering and depth accuracy while achieving competitive tracking accuracy and frame rates

Critical strengths
  • Strong rendering and depth results across indoor and outdoor benchmarks. The proposed method is evaluated across four benchmarks ranging from small indoor scenes (7-Scenes, TUM-D) to mixed clips (RE10K) and large-scale outdoor sequences (Oxford Spires). On the tested sequences, LingGS consistently achieves the best novel-view synthesis quality among the evaluated baselines and yields the lowest L1 rendered-depth error while maintaining a fast processing rate.
  • The proposed decoupling strategy is broadly applicable across different feed-forward architectures. Replacing each model's pose estimation head with the explicit geometry solver across four distinct streaming foundation models (XStreamVGGT, STream3R, TTT3R, and LingBot-Map) consistently improves tracking, reducing ATE with modest computational overhead.
Critical weaknesses
  • The methodological novelty is limited and the technical contribution is incremental. The proposed framework is largely an incremental combination of established components, a frozen pre-trained streaming foundation model (LingBot-Map) used as a virtual depth sensor, a classical feature-based tracking frontend (ORB-SLAM3) extended with a direct depth residual, and keyframe-replay Gaussian optimization, which is borrowed from OTF. The theoretical contribution is minimal, the mathematical re-formulation of the direct depth residual in Appendix A is a standard relative-uncertainty weighting for direct depth residuals rather than a fundamental algorithmic breakthrough. The system behaves more like a polished engineering integration than a novel SLAM methodology.
  • The evaluation is conducted on custom subsets of benchmark datasets rather than standardized splits—specifically, 60 selected short clips out of the vast RE10K dataset, 11 handheld sequences from TUM-D, and 6 clips from Oxford Spires—without disclosing the exact sequence IDs or selection criteria for RE10K and TUM-D. Given that ORB feature tracking is known to struggle in textureless or fast-motion scenarios (as seen with Photo-SLAM failing on 7-Scenes 'Pumpkin', and acknowledged by the authors in Conclusion, there is a strong concern regarding selection bias toward scenes with rich visual features and smooth camera motion, e.g., potentially excluding textureless TUM-D fr3_nostructure sequences.
  • The pipeline is limited to local map tracking and local bundle adjustment, with no loop closure, global pose optimization, or map deformation to correct accumulated drift on long or looping trajectories. In addition, globally consistent monocular GS-SLAM baselines with learned dense trackers, such as Splat-SLAM and HI-SLAM2, are cited in Related Work but left out of Table 1, even though HI-SLAM2 achieves real-time frame rates alongside online loop closure. Without comparing against these methods, the practical advantage of the proposed local pipeline remains unclear.
Questions for authors
  • Could the authors clarify the exact selection protocol/criteria used for picking the 60 clips from RE10K and 11 sequences from TUM-D? What are the exact sequence IDs for the tested datasets? Were these selected randomly, or based on specific motion/visual characteristics? Providing evaluation results on standard benchmark splits would help address potential selection bias concerns. In addition, how does LingGS perform on low-texture or fast-motion sequences, e.g., TUM-D fr3_nostructure sequences, where ORB feature extraction and matching typically struggle?
  • Why were globally consistent monocular GS-SLAM baselines with learned dense trackers, such as HI-SLAM2 and Splat-SLAM, omitted from Table 1? Crucially, HI-SLAM2 runs at real-time frame rates with online loop closure. Could the authors provide a quantitative comparison on 7-Scenes or TUM-D and discuss how LingGS behaves on loopy trajectories without loop closure or map deformation?
  • Could the authors clarify what specific algorithmic or theoretical components in LingGS are genuinely novel beyond the system-level integration of existing tools (LingBot-Map, ORB-SLAM3, OTF) and the standard relative-uncertainty weighting?
Reviewer_R4jJOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents LingGS, an online monocular Gaussian splatting SLAM system. A frozen streaming 3D foundation model (LingBot-Map) acts as a virtual RGB-D sensor, providing intrinsics and per-frame depth. Camera poses come from an ORB-SLAM3-based geometric tracker, not from the model’s camera head. The tracker replaces ORB-SLAM3’s virtual-stereo residual with a direct depth residual under a constant relative-uncertainty model. Mapping follows an OTF-style incremental Gaussian mapper with Laplacian-guided insertion, keyframe replay and an inverse-depth loss. Experiments on 7-Scenes, Oxford Spires, RE10K and TUM-D report better rendering quality and depth accuracy than MonoGS, S3PO-GS, Photo-SLAM and OTF, at roughly 5–8 FPS. The paper also shows that the geometric tracker outperforms the native camera heads of several streaming backbones.

Critical strengths

The system is practical and complete. It requires no retraining and runs at near-interactive rates on a single GPU. The finding that explicit geometric tracking on top of predicted depth beats learned camera heads across four streaming backbones (Table 4) is useful for the community and for future streaming model design. Rendering and depth results are consistently strong against the chosen baselines. The appendix analysis of virtual-stereo versus direct-depth weighting is clear.

Critical weaknesses
  1. Baselines are too weak. All compared methods either use no learned dense depth prior (MonoGS, Photo-SLAM) or target a different setting (OTF, S3PO-GS). The most relevant category is learning-based monocular SLAM with dense depth priors. Examples are DROID-Splat, Splat-SLAM and HI-SLAM2, all discussed in the related work but never compared against. Please try to at least add a comparison with DROID-Splat. This would affect my ratings.

  2. Limited novelty and missing ablations to justify the system design. Technical novelty is limited. The tracker is ORB-SLAM3 with one modified residual, and the mapper largely follows OTF. I think it is acceptable for a solid system paper in SLAM to have limited novelty. However, a system paper then needs to justify its design choices through careful ablations, and several key ones are missing:

    • Why LingBot-Map depth? The paper never ablates the depth source inside the full system. Would the same pipeline work with per-frame monocular depth (e.g., Metric3D, Depth Anything, UniDepth), as used in DROID-Splat? Would it work with depth from the other streaming backbones in Table 4? Table 4 only reports tracking on a single sequence, not end-to-end reconstruction.
    • Camera head + BA. The comparison between the camera head (D) and the geometric solver is somewhat unbalanced. One is per-frame regression, the other a full tracker with local map and local BA. A natural and important baseline is to initialize from the camera head and refine with BA or pose-only optimization using the same depth residual. Without this, the claim that learned camera prediction is fundamentally worse than geometric tracking is not convincingly supported.
    • The gain from the proposed direct-depth residual itself is small. In Table 3, virtual stereo (A) and the proposed solver give identical PSNR (26.35), and ATE only drops from 79.5 to 73.8 cm. This suggests most of the improvement over Photo-SLAM comes from the depth prior rather than the proposed tracker modification.

    Without these ablations, I am not convinced that the current configuration is a well-founded system rather than one working combination.

  3. Speed claims are inconsistent. The abstract claims 2–3× the throughput of evaluated prior SLAM systems. However, in Table 1, OTF is faster on 7-Scenes (7.57 vs. 6.99 FPS) and Oxford Spires (8.24 vs. 6.40 FPS). On TUM-D, the proposed method (5.01 FPS) is slower than both Photo-SLAM (5.99) and OTF (6.54). The 2–3× claim only holds against MonoGS and S3PO-GS. The ~7 FPS claim also does not hold on TUM-D. Please revise these claims and clarify whether backbone inference time is included in the reported tracking times; Table 4 shows an identical 16.1 ms for all backbones.

Questions for authors

How does the full system perform with per-frame monocular depth networks instead of LingBot-Map? How does camera head + BA refinement compare with the proposed solver in accuracy and speed? How does the system compare with DROID-Splat on the same sequences? Please clarify the timing breakdown: backbone inference, tracking and mapping.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.